Expected whole-horizon loss of Algorithm 3
DefinitionFirstOrderFG_Supplement_Alg3_expLossbandit-algorithmsgraph-theorymachine-learningp2o-gran-per-chapterp2o-plan-paperp2o-v1
Algorithm 3's expected incurred loss is
evaluated over the finite tree generated by its history-dependent next-arm law.
Definition code
import Mathlib
import Definitions.Def_FirstOrderFG_Supplement_Alg3_play
import Definitions.Def_FirstOrderFG_treeExp
import Definitions.Def_FirstOrderFG_pathLoss
set_option autoImplicit false
namespace FirstOrderFG.Supplement
/-- **Algorithm 3's whole-horizon expected incurred loss.**
Zhang–Zhang–Luo–Mineiro, arXiv:2610.40861, §2, p.3, and Algorithm 3, p.15.
The finite tree samples exactly the arm law after every possible history. -/
noncomputable def Alg3.expLoss {K : ℕ} (E : Fin K → Fin K → Prop) [DecidableRel E]
(T : ℕ) (ℓ : ℕ → Fin K → ℝ) : ℝ :=
FirstOrderFG.treeExp (Alg3.play E T ℓ) T [] (FirstOrderFG.pathLoss ℓ 1)
end FirstOrderFG.Supplement
Source
Zhang–Zhang–Luo–Mineiro, Near-Optimal First-Order Regret for Bandits with Feedback Graphs, arXiv:2610.40861, §2, p.3; Algorithm 3, p.15